Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding

Parallel diffusion decoding can accelerate diffusion language model inference by unmasking multiple tokens per step, but aggressive parallelism often harms quality. Revocable decoding mitigates this by rechecking earlier tokens, yet we observe that existing verification schemes frequently trigger flip-flop oscillations, where tokens are remasked and later restored unchanged. This behaviour slows inference in two ways: remasking verified positions weakens the conditioning context for parallel drafting, and repeated remask cycles consume the revision budget with little net progress. We propose COVER (Cache Override Verification for Efficient Revision), which performs leave-one-out verification and stable drafting within a single forward pass. COVER constructs two attention views via KV cache override: selected seeds are masked for verification, while their cached key value states are injected for all other queries to preserve contextual information, with a closed form diagonal correction preventing self leakage at the seed positions. COVER further prioritises seeds using a stability aware score that balances uncertainty, downstream influence, and cache drift, and it adapts the number of verified seeds per step. Across benchmarks, COVER markedly reduces unnecessary revisions and yields faster decoding while preserving output quality.

Many AI systems generate text one word at a time, which can make them slow. A newer type of language model can work more like filling in a paragraph with blanks: it can fill in many missing words at once, then revise them over several rounds. This has the potential to make text generation much faster, but current revision methods often waste effort. In particular, the model may erase a word and later put back exactly the same word. We call this repeated erase-and-restore pattern a “flip-flop.” This paper introduces COVER, a method that helps these models check their earlier choices more efficiently. Instead of fully removing a word while checking it, COVER keeps useful surrounding information available, so the model can continue making progress on the rest of the answer. COVER also chooses which words are most worth checking, rather than revising too many words unnecessarily. On math and programming tasks, COVER reduces unnecessary revisions and allows these models to produce answers in fewer rounds and less time, while maintaining similar or better answer quality. This makes this style of language model more practical without retraining the model.